Was anyone else surprised that they used simple k-means for clustering the phonemes? I skimmed the paper and it looks like they use a really high k (128), presu

Was anyone else surprised that they used simple k-means for clustering the phonemes? I skimmed the paper and it looks like they use a really high k (128), presumably to avoid having to do an elbow-plot like approach. Maybe the computational benefits of such a simple algorithm outweighed the potentially bad clusters. Thoughts? I'll try to read the paper in more depth later in case they explain the choice and I missed it.